Step 4: Your First Run: The 10-Minute Ollama Route

Updated 2026-09-11 · Verified 2026-09-11

The first three steps were theory. This one is real: ten minutes from now, there will be an AI on your computer that chats with you — offline, free, no account required. Our tool is Ollama, the simplest local model runner around. Install it, and it handles everything else.

Minute 1: Install Ollama

Go to ollama.com and hit Download:

• Mac / Windows: download the installer, double-click, click through — like any other app.
• Linux: open a terminal and paste one line:

curl -fsSL https://ollama.com/install.sh | sh

When it’s done, open a terminal (Terminal on Mac, PowerShell or Command Prompt on Windows) and type:

ollama --version

See a version number? You’re in.

Minutes 2–8: One command to run it

In the terminal, type:

ollama run llama3.2

llama3.2 defaults to the 3B version — the starter tier from the last step that runs on almost any computer. The first run downloads the model automatically: about 2 GB, a few minutes depending on your connection. When you see the >>> prompt, the model is already running. Just type and chat.

Ask it something — “explain black holes in three sentences.” Watch the words appear one by one: that’s your computer thinking on its own, writing token by token.

A few in-session commands (they start with a slash):

• /bye — exit the conversation.
• To chat again later, run ollama run llama3.2 once more — no re-download needed.

Two things worth knowing while you play. First, the model stays loaded in memory for a few minutes after you exit, so your next session starts instantly; it unloads itself when idle. Second, everything you just did is reversible: ollama rm llama3.2 deletes the model and frees the disk space, and nothing else on your system is touched. Feel free to experiment — ask it to summarize a paragraph, explain a concept, or draft an email. Notice how it responds word by word rather than all at once; that streaming pace is your hardware working, and it’s exactly what the speed numbers in the next minutes are about.

Minute 9: See how much it’s eating

While the model is running, open a second terminal window and type:

ollama ps

You’ll see something like:

NAME           SIZE    PROCESSOR
llama3.2:...   4.0 GB  100% GPU

Two columns worth understanding:

• SIZE: the model’s current combined RAM+VRAM footprint. The file was only 2 GB — why ~4 GB here? Recall the last step: the file is roughly the weights; a running model also needs its “conversation memory” plus runtime overhead (theoretical estimate: ~4.4 GB total at 8K context; Ollama’s default context is shorter, so real-world numbers are usually smaller).
• PROCESSOR: 100% GPU means the model lives entirely on your graphics card — the ideal state. A ratio like 50%/50% CPU/GPU means VRAM ran out and part of the model spilled into system memory. It still runs, just noticeably slower.

If you have an NVIDIA GPU, open another window and run nvidia-smi to see VRAM usage and utilization; on Mac, open Activity Monitor and watch memory pressure.

Minute 10: Is the speed normal?

No GPU at all? It still works — speed is then limited by memory bandwidth, and a 3B model on pure CPU manages a few dozen words per second (theoretical estimate, ±30%), which is perfectly usable for chat. With a GPU it’s much faster: an RTX 3060 theoretically tops 100 words per second. If output crawls at one or two words per second, the model almost certainly didn’t fit in VRAM — go back and check the PROCESSOR column in ollama ps.

Later, when you want full control over every knob (GPU layers, context length, GGUF files you downloaded yourself), you can graduate to llama.cpp — the underlying engine Ollama actually runs on. For the full comparison of both paths and every parameter, read Running GGUF Locally: llama.cpp and Ollama.

That’s the zero-to-one leap, done. Don’t stop at reading — go do it now: